A Multitask Learning Model with Multiperspective Attention and Its Application in Recommendation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• E-commerce Analytics

Background:

• Multitask learning is crucial for e-commerce, aiming to predict user clicks and orders simultaneously for improved user satisfaction and business outcomes.
• Existing methods often model user representation from historical behavior sequences, but user interests can dynamically change.
• Capturing evolving user interests is challenging for traditional models.

Purpose of the Study:

• To introduce multi-perspective attention and sequential behavior into multitask learning for a deeper understanding of user interests and decision-making.
• To enhance parameter sharing flexibility and preserve task-specific feature advantages within multitask learning frameworks.

Main Methods:

• Proposed a novel multitask learning framework incorporating multi-perspective attention mechanisms.
• Introduced an improved attention mechanism focusing on expert interaction for flexible parameter sharing.
• Pioneered four interaction modes in multitask learning: implicit, explicit hard, explicit soft, and data fusion.

Main Results:

• The proposed model demonstrated a superior understanding of user interests and decision-making processes.
• Achieved significant improvements over state-of-the-art methods on both public and lab medical datasets.
• Consistently outperformed existing approaches in predicting click and order targets.

Conclusions:

• The integration of multi-perspective attention and sequential behavior significantly enhances multitask learning for e-commerce applications.
• The novel interaction modes offer flexible and effective parameter sharing strategies.
• The model provides a robust solution for understanding dynamic user interests and improving prediction accuracy.